DeepSeek folds vision into pretraining: V4.1 Flash snapshot ships with KV cache cut to a quarter of the prior generation
Synopsis
DeepSeek released the October 17, 2026 snapshot of V4.1 Flash, served under the deepseek-flash API model name. It is the company's first sparse mixture-of-experts model built on its Causal Encoder-Decoder architecture, activating 8B parameters on input and 16B on output from a 552B backbone, with image understanding trained jointly with text from the start of pretraining rather than added afterward. Compressed KV caching cuts cache memory to roughly a quarter of the previous Flash generation, and DeepSeek reports it exceeds V4 Pro on performance, speed and task completion time.
Interpretation
DeepSeek V4.1 Flash 1017 is the October 17, 2026 snapshot of V4.1 Flash, served by DeepSeek under the deepseek-flash API model name and positioned as the cost-efficient tier of the V4.1 family. The earlier experimental V4 Flash Vision Exp added image capability after training, rather than training visual and text embeddings together from the outset. Image understanding is native to the architecture, with visual and text embeddings trained jointly from the start of pretraining; DeepSeek also reports the model exceeds V4 Pro on performance, speed and task completion time.
The model uses a sparse mixture-of-experts structure, so each token calls only part of the parameters and per-token compute stays low relative to the 552B total size. A sparse mixture-of-experts model is composed of many expert subnetworks, and each forward pass activates only a few of them, decoupling compute from total parameter count. From a 552B-parameter backbone it activates 8B parameters on input and 16B on output, an asymmetric split that keeps per-token compute low.
Compressed KV caching cuts cache memory to roughly a quarter of the previous Flash generation, significantly reducing costs on agentic workloads. KV cache refers to the stored key and value vectors for past tokens during autoregressive generation, which avoids recomputation but grows with context and is a major cost in long tasks. The model targets coding, terminal and computer-use agents plus long-horizon tasks and long-context analysis; OpenRouter reports 99.94% availability over the last 3 days, above 97.18% without routing.
Perspective
This report draws on a model-page summary covering positioning, architecture, parameter activation, caching and availability data, not a full technical report. It is intended for engineers evaluating API choices and researchers following sparse mixture-of-experts architectures. The material states no follow-up experiments or release plans.
The conclusion that it exceeds V4 Pro on performance, speed and task completion time is reported by DeepSeek, and the material lists no corresponding benchmarks or scores. The availability figures are OpenRouter monitoring over the last 3 days, reflecting hosting-side behavior rather than model capability.